Program generation for schema-based, typed data access

Transcription

1 Program generation for schema-based, typed data access Ralf Lämmel Software Engineer Facebook, London

2 Program generation A use case at Facebook Purpose of generation: typed data access ("O/R mapping" et al.) Meta = Object language = Hack (not to be confused with PHP) Input for generation: schemas (or mapping definitions)

3 Generate access code from schema We illustrate the use case with opensource software.

4 The underlying codegen library We illustrate the use case with opensource software.

5 Signing of generated code Key ideas: Generate hash code from generated code. Prevent diffs that change the generated code. We illustrate the use case with opensource software.

6 Further reading Introducing "hack-codegen" Open source code base for hack-codegen Typed data access proof of concept (some form of EntSchema) Docs on hack-codegen Information on the "Ent" framework This is publicly available information which does not perfectly match Facebook's internal approach.

7 Program generation challenges not just necessarily within the chosen scope at Facebook Metaprogramming support? Customization of generated code? Integration of generation with development? Complexity of generator! Size of the generated code!... Let's look at these challenges, one by one, in terms of the aforementioned use case and its Facebook instance.

8 Metaprogramming support? Staging? Compile- or runtime generation may be too slow. It may disrupt development (e.g., source-code-based tooling). Staged programming support was not there to start with. Type safety? The generators are tested a lot. Concrete object syntax? Code construction is served by fluent APIs. There is a lot of logic; there is less templates.

9 Customization of generated code Manual sections are removed from signature for signing. Signing helps with avoiding unintended changes. Manual sections are kept when re-generating code. - We may need to manage partially generated code. - Changes of defaults may be non-obvious. - Customized code may fail to break when it should. - Host language may limit factoring of customization. We illustrate the use case with opensource software.

10 Integration of program generation with development + We generate code to test generated code. - Generated code may be out of sync with database. - Generated code may permit too much on database. - Generated code may swamp version control. - Generated code is in the face of the developer.

11 Complexity of generator - Developers may face many different files per schema. - The code generator is tied into lifecycle management. - The codebase may get locked into large API. - The evolution of the generator may be difficult.

12 Size of the generated code - We start to have a problem, when the size of generated code is in the same ballpark as the size of the genuine code. Think of the time to... re-generate, teach IDE, compile, run tests, commit,...

13 Wanted! Improvements: Generate less code! Compile less code! Version-control much less code! Don't think so much in terms of code! Better separate generated and non-generated code! Preservation: Data access is typed! Get in touch, if interested. Provide IDE support such as code completion! Enable intuitive customization of generated code!

14 Towards a relatively systematic literature survey on program generation

15 Questions of interest Broad category Does the approach address a case study (on program generation), or language support (for program generation), or an application domain (of program generation), or something else? In reality, approaches may combine all of these.

16 Questions of interest Application domain If the approach addresses a case study, then what is the underlying application domain? If the approach addresses language support, then does it point at some application domain? If the approach addresses an application domain, then which one is it and how well does it relate to data access?

17 Questions of interest Language aspects What is the object language? What is the metalanguage? What language support is facilitated? What infrastructural support (e.g., a library) is facilitated? In reality, meta = object language may be a common case, of course.

18 Questions of interest Generation time Is the code generated at... (RT) run time or (CT) compile time or (BT) build time (i.e., ahead of compile time)? Does the generated code provide guarantees... at run time (e.g., typed data access) or at compile time (e.g., well-typedness)?

19 Questions of interest Customization Is customization of generated code facilitated, and, if so,... is customization achieved by some form of specialization (e.g., subclassing) or some form of code changes, and, if so,... is round-tripping / synchronization supported, and, if so,... by what means and to what extent, or by other means and is such customization essential, and, if so, because of which specifics (language or domain)?

20 Questions of interest Software engineering How is code generation addressed by the software building? How is code generation addressed by software testing? How is code generation addressed within the IDE? What other software engineering aspects are addressed?

21 Questions of interest System design Does the generated code entail... a black/gray box (e.g., a parser)... as a service or at the command line or a library (essentially an interface) or a framework (typically requiring specialization) or a template (typically requiring some code changes)?

22 Venues GPCE DSL SAIG GTTSE SLE PEPM ICFP POPL PLDI

23 GPCE 2017 papers [...] Case study Language support Application domain Software engineering Time Customization Typed data access HSpiral Search CT CaVa Multi-lingual & distributed BT Guix RT PAX Synthesis BT QSR MetaML Library optimization CT Formal semantics RT Haskino Debugging BT Silverchain SpiralS Usability of APIs High performance BT BT Control Type soindness RT Tensor Reusability BT

24 [HSpiral] A Haskell Compiler for Signal Transforms (GPCE'17) Domain: signal transforms (FFT) Meta: Haskell Object: Haskell, C, deep embedded DSLs Concepts: type classes, monads, GADTs, type families, index types, rewriting, compilation (to C), deep embedding, quasi-quotation, search (for optimization)

25 [CaVa] Automatic Generation of Virtual Learning Spaces Driven by CaVaDSL: An Experience Report (GPCE'17) Domain: virtual learning spaces for cultural heritage Meta: Java Object: scripts, SPARQL, HTML, CSS, PHP,... Concepts: external DSL, multi-lingual and distributed target

26 [Guix] Code Staging in GNU Guix (GPCE'17) Domain: OS configuration and system administration Meta: Scheme Object: embedded DSLs for package definitions and build actions Concepts: staging, macros, S-expressions

27 [PAX] Parser Generation by Example for Legacy Pattern Languages (GPCE'17) Domain: parsing for legacy languages Meta: C# (for metamodel inference and C# generation) Object: metamodels (for patterns), C# Concepts: pattern synthesis (grammar inference), finite state machines

28 [QSR] Quoted Staged Rewriting: A Practical Approach to Library-De ned Optimizations (GPCE'17) Domain: stream fusion Meta: Scala Object: Scala Concepts: staging, quasi-quotation, rewriting, macros, CPS, inlining

29 [MetaML] Refining Semantics for Multi-stage Programming (GPCE'17) Domain: None (run-time code generation by staging) Meta: MetaML Object: ML Concepts: explicit substitutions, structural operational semantics

30 [Haskino] Rewriting a Shallow DSL using a GHC Compiler Extension (GPCE'17) Domain: embedded systems / microcontrollers Meta: Haskell Object: C Concepts: shallow and deep embedding, debugging, firmware interpretation, translation, monads, shallow-todeep transformation, compiler plugin

31 [Silverchain] Silverchain: A Fluent API Generator (GPCE'17) Domain: fluent API generation Meta: not defined (Java?) Object: grammars (input), class definitions (output) Concepts: fluent APIs, deterministic push-down automata, BNF

32 [SpiralS] Staging for Generic Programming in Space and Time (GPCE'17) Domain: convolution on images and FFT Meta: Scala Object: Java Concepts: lightweight modular staging (LMS), polymorphism, generic programming

33 [Control] Staging with Control: Type-Safe Multi-stage Programming with Control Operators (GPCE'17) Domain: None (run-time code generation by staging) Meta: None (a calculus in MetaML style) Object: None Concepts: staging, control operators (shift0, reset0), computational effects, type soundness, type inference

34 [Tensor] Towards Compositional and Generative Tensor Optimizations (GPCE'17) Domain: Tensor computations (quantum chemistry and physics, big data analysis, machine learning, computational fluid dynamics) Meta: Undefined Object: C, custom intermediate language Concepts: intermediate language

35 Thanks!